Compatibility

Package Line

The 0.23.x harness line requires Python >=3.12,<3.15 and robotics-runtime-contracts>=0.22,<0.23. The 0.22.x line requires the same Python range and robotics-runtime-contracts>=0.21,<0.22.

The changelog and release procedure record exact release prerequisites. Installed package metadata, tagged source and distribution digests identify the selected pair. Package support and native profile qualification remain separate.

Each role retains its canonical default. Datasets use v2 for one or multiple MCAP members. Explicit v2 scenario/runtime/observation/result/qualification roles support the native archive example and distinct archive assessments. Historical datasets require their matching archived package; readers do not convert old forms.

Compatibility Tuple

A qualification composition is compatible as one tuple, not as independent latest versions: the CLI package pair (robotics-runtime-contracts and robotics-acceptance-harness), the evaluator artifacts (wheel or OCI image) pinned by exact SHA-256 digests, the document schema versions they emit and consume, and the operator trust policy that admits those evaluators. Replacing one member without retesting the whole tuple is unsupported.

Select one tested tuple and retain its package pins, evaluator digests, schemas and trust policy together. Infra images and native execution profiles have their own qualified digests; a Python package release does not update them.

The first-result recipe uses this offline tuple:

Member

Selection

CLI packages

Exact installation command in the first-result recipe; source example

Evaluator

Built-in OTLP metric method; no external evaluator wheel or OCI image

Image digest

None; this tuple executes no container image

Schemas

Catalogue-selected v1 scenario, runtime, run, evidence and result documents

Trust policy

None required; no external evaluator is admitted

Example producer

MCAP writer, pinned in the first-result installation command

The example deliberately returns incomplete for the missing live observations. Its guide explains the validation, evaluation and exit codes. External evaluator wheels, OCI images and their operator profiles form separate tuples; qualify those exact artifacts before use.

Document Set

The authoritative role-to-schema mapping is the contracts package catalog. The harness calls its public schema_for_role(), schema_versions_for_role() and validate_role() APIs and maintains no second compatibility table. schema_for_role(role) preserves the canonical default; schema_for_role(role, version=2) selects an explicit supported version. Declared supported versions are validated without converting input bytes.

Unknown schema versions, wrong document roles, and contradictory bundle facts fail before observation or evaluation begins. Scenario extensions remain separately versioned and digest-pinned by their canonical URI.

The caller’s pinned extension registry is forwarded to every loaded bundle document, including scenario, runtime, model, dataset, permit and verification. Unknown namespaces, wrong schema digests and invalid payloads fail validation.

Provider Compatibility

The stable interface contains observed capabilities and implementation bindings, not a closed simulator, middleware, storage, or accelerator list. Provider qualification belongs to runtime infrastructure. A new provider is compatible when it emits the existing canonical documents and passes the same conformance suite.

A current package install or source fixture does not qualify an infrastructure composition or native profile.

The Python commands install without ROS. V1 document formats retain required ROS graph, lifecycle and timing fields; live observation uses the ROS 2 packages and interfaces declared by the runtime. V2 evaluates captured native software or simulation observations without constructing ROS fields. It does not launch SDK operations. V2 physical qualification and v1 transport evaluation of v2 profiles are refused.

V2 results separate original run/source/evidence hashes from method configuration, actual software environment and UTC assessment time. Captured source measurement windows and assessment windows use explicit clock identities and integer bounds. Nanosecond document integers are exact on the Python file path. JavaScript bridges must preserve exact files/text, and pass CLI window arguments as decimal strings or bigint. JSON.parse/re-emission through JavaScript Number is outside that exact-byte path. The host rejects numeric window arguments; read-only MCP operations forward file snapshots and keep worker payloads as text. The built-in OTLP method accepts unix_ns; selected product assessments may use native_ns without an inferred UTC conversion. Empty criteria or entirely absent measurement coverage cannot produce a passed verdict. Acceptance-run.v1 and acceptance-aggregate.v1 retain their existing document semantics; separate assessment cohorts use separate aggregates. Explicit assessment controls and qualification-bundle.v2 add the archive assessment path, retaining the original result separately from each new assessment. Canonical qualification-bundle.v1 defaults and historical signed readers are unchanged.

Exact provider and hardware support belongs to a qualified runtime artifact; installing the package or selecting an SDK binding does not establish it.

Dependency Reproducibility

The root uv.lock pins the workspace’s development dependency graph. The contracts workspace source override applies only to development. Releases use uv build --package robotics-acceptance-harness --no-sources and standard index requirements, then test wheel and source distributions outside the workspace with the exact released contracts version from PyPI. See the workspace release procedure for the source, metadata and provenance checks.